Log in Sign up
Productivity Nirva Raises 33 Million Dollars for an 8-Gram AI Pendant That Listens Only to You Chatbots NVIDIA Ships Nemotron 3.5 Bolt E4B: an 8B Open Agent Model That Fits in About 4GB Claude Anthropic Now Operates Claude Directly on Google Cloud With IAM, Marketplace Billing and Opus 5.5 Fast Mode Security UK Puts 100 Million Pounds Into an AI Biosecurity Lab Backed by OpenAI, Anthropic and DeepMind Security Moonshot Stripped Cyberattack Data From Kimi K3.1 Pro's Training to Cut Misuse Risk Gemini Gemini 4 Argon Processed Nearly 25 Trillion Tokens in Its First 72 Hours — Triple Gemini 3's Pace Business Jensen Huang Says a Memory Glut by 2029 Is Unlikely as the AI Buildout Keeps Swallowing HBM and LPDDR Business AI Film Studio Utopai Raises $60 Million at a $1 Billion Valuation With More Than 20 Projects in Production Productivity Nirva Raises 33 Million Dollars for an 8-Gram AI Pendant That Listens Only to You Chatbots NVIDIA Ships Nemotron 3.5 Bolt E4B: an 8B Open Agent Model That Fits in About 4GB Claude Anthropic Now Operates Claude Directly on Google Cloud With IAM, Marketplace Billing and Opus 5.5 Fast Mode Security UK Puts 100 Million Pounds Into an AI Biosecurity Lab Backed by OpenAI, Anthropic and DeepMind Security Moonshot Stripped Cyberattack Data From Kimi K3.1 Pro's Training to Cut Misuse Risk Gemini Gemini 4 Argon Processed Nearly 25 Trillion Tokens in Its First 72 Hours — Triple Gemini 3's Pace Business Jensen Huang Says a Memory Glut by 2029 Is Unlikely as the AI Buildout Keeps Swallowing HBM and LPDDR Business AI Film Studio Utopai Raises $60 Million at a $1 Billion Valuation With More Than 20 Projects in Production

NVIDIA Ships Nemotron 3.5 Bolt E4B: an 8B Open Agent Model That Fits in About 4GB

Chatbots · By: AIReport Admin · 1 hour ago · 👁 10
NVIDIA Ships Nemotron 3.5 Bolt E4B: an 8B Open Agent Model That Fits in About 4GB

NVIDIA released Nemotron 3.5 Bolt E4B 4GB on October 23, an 8-billion-parameter model with 4 billion effective parameters, post-trained from Google Gemma 4 E4B-it for local voice assistants, interactive characters and robot interfaces. NVIDIA reports 3.1 to 5.0GB of VRAM and 52 to 133 tokens per second on RTX and Jetson hardware; a 6.18GB Q4 GGUF runs on machines with 6GB of RAM.

NVIDIA has released Nemotron 3.5 Bolt E4B 4GB, a small open-weight language model built for one specific job: being the conversational brain of local voice assistants, in-game characters and robot front ends. The model went generally available on Hugging Face and build.nvidia.com on October 23, a date NVIDIA had committed to in an October 15 blog post.

The specs are deliberately modest. Bolt E4B is post-trained from Google's Gemma 4 E4B-it and carries 8 billion total parameters with roughly 4 billion effective. It handles text in and text out with up to 64,000 tokens of context in each direction, and it ships with controls that matter for speech products: a thinking mode that can be switched on or off, a runtime thinking budget, and a conversational mode tuned for concise spoken-style replies.

Where it runs is the story. NVIDIA's own measurements report peak VRAM of 3.1 to 5.0GB across RTX and Jetson configurations, with 52 to 133 tokens per second, and the company pitches the model for GeForce RTX, Jetson, RTX Spark and DGX Spark hardware. A 6.18GB Q4_K_M GGUF plus a BF16 safetensors version are on Hugging Face, and the quantized build runs on machines with around 6GB of RAM. An optional 737MB DFlash drafter speeds up generation further. NVIDIA's accuracy figures come from its NeMo Gym harness and should be treated as vendor-reported.

On benchmarks, NVIDIA claims roughly 40 percent higher average scores than its earlier Nano 4B model on its selected tests, though the company itself cautions that benchmarks only partly predict how an agent behaves in a real application. The company also demonstrated the model on a Reachy Mini desktop robot, where Bolt handles spoken responses and triggers animation scripts while a separate Cosmos 3 Edge model supplies vision, all running locally on a Jetson Orin Nano 8GB.

The release follows the Nemotron 3.5 Bolt 74B, a 9-billion-active-parameter model in a 24GB footprint that NVIDIA shipped earlier in October. Together the two define a bracket: the 74B for capable on-prem agents, the E4B for the cheap, always-on, low-power end where the model shares silicon with a microphone and a motor controller. Licensing differs from the base model as well, with Bolt E4B under the OpenMDW-1.1 license on top of Gemma 4's Apache 2.0 base.

The interesting tension is what "4GB" actually buys. A model that fits in a phone or a 249-dollar Jetson kit cannot reason like a frontier system, and NVIDIA is not claiming it does. The bet is that the next wave of agent deployments lives on-device for latency, privacy and cost, and that the model that wins that market is the one that answers in three words instead of three hundred.

Comments (0)

Log in to join the discussion

Log In

No comments yet